christopherheuer/phi3-mini-4k-instruct-pii-redactor

🧠 Phi-3 Mini Instruct Redactor

This is a fine-tuned variant of microsoft/Phi-3-mini-4k-instruct, specialized in document redaction.

The model was fine-tuned using the LoRA method via PEFT and TRL, then merged into the base model to enable direct inference without needing adapter weights.

📚 Training Details

• Base Model: microsoft/Phi-3-mini-4k-instruct
• Tuning Method: LoRA (merged)
• Data Format: JSONL ({"instruction": ..., "output": ...})
• Trainer: SFTTrainer from Hugging Face TRL
• Epochs: 3
• Batch Size: 4 (grad. acc. 2)
• Quantization During Training: 4-bit (bnb nf4)

✅ Final weights are full precision (fp16/fp32), ready for inference.

🚀 Example Usage

from transformers import AutoTokenizer, AutoModelForCausalLM

model = AutoModelForCausalLM.from_pretrained("your-username/phi3-mini-instruct-redactor")
tokenizer = AutoTokenizer.from_pretrained("your-username/phi3-mini-instruct-redactor")

prompt = "Remove all personal data from the following text:\nJohn Doe lives at 123 Elm Street."
inputs = tokenizer(prompt, return_tensors="pt")
outputs = model.generate(**inputs, max_new_tokens=128)
print(tokenizer.decode(outputs[0]))

🧠 Intended Use

• Redacting personal/sensitive information
• Fine-tune starting point for downstream tasks

🧱 GGUF version available?

Yes! You can use this model in llama.cpp or other GGUF-compatible inference tools:

👉 Link to gguf version

📖 Dataset Attribution

This model was fine-tuned using a modified version of the NinjaMasker-PII-Redaction Dataset by King-Harry, which is licensed under the Apache 2.0 License. The dataset was reformatted to suit the training requirements of this model.

⚠️ Disclaimer:

This model is provided "as-is" without any warranties or guarantees regarding its performance, accuracy, or reliability in detecting and redacting personally identifiable information (PII) or other sensitive data.

The model may fail to identify or fully redact all forms of PII, depending on input format, context, or model limitations.

Use of this model is at your own risk.

The authors and maintainers of this model accept no responsibility or liability for any data leakage, compliance violations, or security breaches that may occur as a result of using this model.

📜License

MIT License